Abstract:

With the widespread deployment of face recognition technology in critical fields such as identity authentication and financial payment, the authenticity and security of face data have become core challenges in data security governance. Existing methods, such as those based on latent space reconstruction error that rely on diffusion model priors, insufficiently sensitive to subtle tampering that only affects local regions, as the local forgery error signals are easily diluted by global reconstruction errors. To address this issue, this paper proposed an improved reconstruction error detection method based on an autoencoder, which was distinct from diffusion-model-driven latent space reconstruction error and integrates dual mechanisms of high-frequency feature enhancement and key region protection, specifically optimized for local tampering detection in a lightweight manner. First, generated spatial weight mask via a multi-task cascaded convolutional network to guide the model to focus on key regions such as facial features; then, it designed a lightweight learnable high-pass filter to adaptively extract and enhance high-frequency forgery traces in images; finally, by incorporating both mechanisms into the reconstruction error calculation, it achieved cooperative enhancement of the error signals in both spatial and frequency domains. Experimental results show that the proposed method achieves significant performance improvement in local tampering detection tasks, with a local tampering sensitivity of 89.4%; in cross-dataset tests, it attains average precision of 92.1% on DeepFakeDetection and 90.8% on WildDeepfake, demonstrating good generalization capability; meanwhile, it exhibits only minor performance degradation under noise and compression perturbations, reflecting strong robustness.

Key words: data security governance, face forgery detection, latent space reconstruction error, high-frequency feature enhancement, key region protection